- Email: [email protected]
Wavelet Neural Networks for DNA Sequence Classification Using the Genetic Algorithms and the Least Trimmed Square
Available online at www.sciencedirect.com
ScienceDirect Procedia Computer Science 96 (2016) 418 – 427
Wavelet Neural Networks for DNA Sequence Classification Using the Genetic Algorithms and the Least Trimmed Square Abdesselem Dakhlia*, Wajdi Bellilb, Chokri Ben amarc b
a Department of Computer Science, REGIM, University of Gabes 6002 Gabes, Tunisia Department of Electrical Engineering, REGIM, University of Gafsa 2110 Gafsa, Tunisia c Department of Electrical Engineering, REGIM, University Sfax 3018 Sfax, Tunisia
Abstract This paper presents a structure of the Wavelet Neural Networks used to classify the DNA sequences. The satisfying performance of the Wavelet Neural Networks (WNN) depends on an appropriate determination of the WNN structure optimization problem. In this paper we present a new method to solve this problem based on Genetic Algorithm (GA) and the Least Trimmed Square (LTS). The GA is used to solve the structure and the learning of the WNN and the LTS algorithm is applied to select the important wavelets. First, the scale of the WNN is managed by using the time-frequency locality of wavelet. Furthermore, this optimization problem can be solved efficiently by Genetic Algorithm as well as the LTS method to improve the robustness. The performance of the Wavelet Networks is investigated by detecting the simulating and the real signals in white noise. The main advantage of this method can guarantee the optimal structure of the WNN. The experimental results have indicated that the proposed method (WNN-GA) with the k-means algorithm is more precise than other methods. The proposed method has been able to optimize the wavelet neural network and classify the DNA sequences. Our goal is to construct a predictive approach that is highly accurate results. In fact, our approach allows avoiding the complex problem of form and structure in different groups of organisms. The experimental results are showed that the WNN-GA model outperformed the other models in terms of both the clustering results and the running time. In this study, we present our system which consists of three phases. The first one is the transformation, is composed of two sub steps; the binary codification of the DNA sequences and the Power Spectrum Signal Processing. The second step is the approximation; it is empowered by the use of the Multi Library Wavelet Neural Networks (MLWNN). Finally, the third one is the clustering of the DNA sequences, is realized by applying the algorithm of the k-means algorithm.
© 2016 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license © 2016 The Authors. Published by Elsevier B.V. (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under under responsibility Peer-review responsibilityofofKES KESInternational. International Keywords:Wavelet Neural Networks, GA, DNA sequences, LTS, MLWNN, K-means ;
1. Introduction The WNN have recently attracted extensive attention for its ability to identify effetely nonlinear dynamic systems with incomplete information [1-2-3-4-5]. The WNN were introduced by Zhang and Benveniste in 1992 as combination of the artificial neural network wavelet decomposition [1-2-3]. The generalization performance of the WNN is trained by least-square approach deteriorates when outliers are present. This training approach involves estimating parameters in the network by minimizing some function costs, a measure reflecting the approximation quality is performed by the network over the parameter space in the network. However, the studies of the WNN have often concentrated on small dimension [6]. The reason is that the complexity of network construction will be raised exponentially with the input dimension, i.e. the curse of dimensionality to improve the performance of the WNN in high dimension application. It is a key problem that how to appropriate determinate the network
* Corresponding author. Tel.: +21628002213; fax:+ 216 75 270 686 E-mail address: [email protected]
1877-0509 © 2016 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of KES International doi:10.1016/j.procs.2016.08.088
419
Abdesselem Dakhli et al. / Procedia Computer Science 96 (2016) 418 – 427
structure. This structure has been studied by several researchers. The research effort has been made to deal with this problem over the last decades [6-7-8-9]. The method, which is referred to as Matching Pursuit (MP), was first introduced by Mallat [10]. The Neural Network was used to solve the classification system, such as classification of the DNA sequences using the artificial neural networks [11]. Agnieszka E. et al used method to classify the genomic sequences. This method is combined a wavelet analysis and a self-organizing map algorithm [12]. It is used to extract feature of the oligonucleotide patterns of a sequence. The variation is quantified by the estimated wavelet variance, which yields a feature vector. This method is allowed the results to be visualized. When only thousands of nucleotides are available, wavelet-based feature vectors of short oligonucleotide patterns are more efficient in discrimination than frequency-based feature vectors of long patterns The Wavelet analysis is applied to extract the features of the DNA sequences [13]. The classification of two types of DNA sequence is studied. As well as, 20 samples of the artificial DNA sequences whose their types are known, are given in order to recognize the other types of the DNA sequences. The way of wavelet denoising keeps more main formation in the original curve but it cannot predict accurately the gas zones. In addition, the Wavelet analysis of frequency chaos game signal has been used to classify the DNA sequences. The results stemming from the complex Morlet wavelet analysis of the frequency chaos game signals have presented its accuracy in detection of variable DNA sequences structures. Moreover, this could serve in discovering unknown domains with potential biological significance in genomes [14-15-19]. This paper is organized as follow: in section II, we present an overview of the proposed method. In this section we present the transformation of the DNA sequences and the theory of wavelet neural networks. Section III deals with the simulation results of the proposed approach, which is used to classify the DNA sequences and Section IV ends up with a conclusion and a discussion. 2. Methods This paper presents a new method of clustering of the DNA sequences based on wavelet network using the Multi Library Wavelet Neural Networks (MLWNN) to approximate f(x) of the DNA sequence. The genetic algorithm is used to solve the structure and the learning of the WNN. This approach is divided into two stages: approximation of the input signal and classification of feature extraction of the DNA sequences using the Wavelet Neural Network (WNN) and the k-means algorithm. 2.1. Conversion of the DNA sequence into a genomic signal Our system is used the DNA sequences components to classify the species. These sequences are composed of four basic nucleotides (A, G, C and T) are called adenine, guanine, cytosine and thymine respectively, where each organism is identified by its DNA sequence [20-15-16]. The feature extraction of the DNA sequences can be viewed as finding a set of vectors which represents effectively information [17]. The method of the indicator translates the data into a digital format, which indicate the presence or absence of four nucleotides [18]. The binary indicator sequence is constructed to replace the individual nucleotides with values either 1 or 0. 0 stands for absence and 1 for presence of a particular nucleotide in specified location in DNA signal [19-20]. For example, if x[n] =[T T A T G TA ...], we obtain: x[n] =[0001 0001 1000 0001 0010 0001 1000 . ..] 2.2. Fourier Transform and Power Spectrum Signal Processing After the DNA sequence data have been constructed into these indicator sequences, they can be manipulated with mathematical function. The Fourier Transform is used to each indicator sequence x(n) and a new sequence of complex numbers, named f (x) , is obtained: N 1
¦
f (x)
n
X 0
e
(n )e
jS n / N
,k
0 ,1, 2 , ...N 1
(1)
It is easier to work with sequence Power Spectrum, rather than original discrete Fourier Transform. The Power Spectrum Se[k] is defined as, Se[ k ]
f ( x)
2
(2)
Where the frequencies k= 0, 1, 2, …, N-1. 2.3. Wavelet Neural Network and Time-frequency Analysis The combination of the wavelet transform and the artificial neuron networks defines the concept of the wavelet networks, which applied the mother wavelet functions instead of the traditional sigmoid function as a transfer function of the each neuron. It is composed of three layers, called the input layer, the hidden layer, and the output layer. It has the same structure as the architecture radial function. The salaries of the weighted outputs are added. Each neuron is joined to the other following layer.
420
Abdesselem Dakhli et al. / Procedia Computer Science 96 (2016) 418 – 427
The Wavelet network (Fig. 2) is defined by pondering a set of wavelets dilated and translated from one wavelet function with weight values to approximate a given signal f. Mapping functions to a time-frequency element training set, the overall response of the WNN is
N
^
¦
y
w
§ x bi · N i ¸ ¦ a © ai ¹ k 0
Z i< ¨
x
k
Where f is the regression function and the error term
H
i 1
k
(3) Where (x1, x2,…,xNi ) is the vector of the input, yˆ is the output of the wavelet neural networks, Nw is the number of wavelets, ai is the dilation, bi is the translation, and wi is the weight of the WNN. There, it is often useful to consider, the decomposition of wavelets cleanly, that the output can have a component refine in relation to the variables of coefficients ak (k = 0, 1... Ni) (Fig.2). wavelet occur in family of function and each is defined by translation bi which controls the position of a single function, called the mother wavelet <(x) and dilation ai which controls the scaling parameter. A WNN can be regarded as function approximator, which estimates an unknown function mapping:
f x H ,
y
(4) is a zero-mean random variable of disturbance.
2.4. Multi Library Wavelet Neural Network(MLWNN) The WNN is constructed by many approaches [2-3], Zhang applied two stages to construct the Wavelet neural Network: First, the discretely dilated and the translated version of wavelet mother function < are used to construct a wavelet library W:
®\ i :\ i ¯
W
x
½
D i\ ai x bi ,D i ¨ ¦ >\ ai xk bi @2 ¸ 2 , i 1,...L ¾, §
©k
·
n
1
¹
¿ (5) Where xk is the sampled input and L is the number of wavelets in W, then the best M wavelets is selected based on the training data from the library W, in order to construct the regression: 1
f x y ¦ w\ x , M
i
iI
i
(6)
Where I is a M-element subset wavelet from the wavelet library, and M d L. Secondly, the minimized of the cost function:
1 § · min ¦ ¨ y ¦ w\ x ¸ ¹ wi , iI n © n
j I
k
k 1
iI
i
i
k
2
(7)
,
Two heuristic algorithms are derived by Zhang, called, the stepwise selection by orthogonalization is used to select appropriate wavelet in the hidden units while the second one is used to choose the number of the hidden units and of the wavelets M, which are selected as the minimum of the so-called Akaike’s final prediction error criterion (FPE)[2-3]:
j
FPE
§ ¨ ¨ ©
f
· ¸ ¸ ¹
1
n
pa
1
n
pa
n
§ 1 n ¨ ¦¨ 2n k 1 ¨ f ©
x k y
· ¸2 ¸ k¸ ¹
n (8) Where npa is the number of the parameters in the estimator. The gradient algorithms are used to train the WNN, while the least mean squares are used to reduce the mean-squared error:
j w
· 1 n §¨ w ¸ 2 , ¦ ¸ n i 1 ¨© y i y ¹
(9) Where j is the real output from a trained WNN at the fixed weight vector w. In this study, the hidden layer is composed with wavelet functions which are selected by using the Least Trimmed Square (LTS) method, which is used for initializing the wavelet neural networks. The residual (or error) ei at the ith ouput of the WNN due to the ith example is defined by :
ei
y i yˆ i , i n
421
Abdesselem Dakhli et al. / Procedia Computer Science 96 (2016) 418 – 427
(10) The Least Trimmed Square estimator is used to select the WNN weights that minimize the total sum of trimmed squared errors :
E
total
1 2
p
l
¦ ¦ e k
1
i
2
(11)
ik
1
The notation for the ordered absolute residuals is e(i) , so
e 1 d e 2 d ... d e n .
2.5. Wavelet Network construction using Genetic Algorithm (GA) 2.5.1 Genetic Algorithm (GA) John Holland presented the Genetic Algorithm (GA) in 1975, is a computing search technique used in finding a solution in the optimization problems. GA applies the principles of evolution found in nature to the problem of finding an optimal solution [2122-23]. GA generates successive populations of the alternate solutions that are represented by a chromosome, i.e. a solution to the problem, until acceptable results are obtained. Each chromosome, individual solution, consists of number of binary code called genes. A fitness function assesses the quality of a solution in the evaluation step. The evolution from one generation to the next is performed using three operations: reproduction, crossover and mutation. The evaluation of the fitness value is used by the chromosomes for reproduction. The fitter chromosomes have higher priority to be chosen into the recombination pool according to the roulette wheel or the tournament selection methods. Crossover selects genes from two parent chromosomes using randomly chosen crossover point and creates two new off springs and mutation process changes chromosome randomly by altering a single bit as in Fig.1.
Figure 1 Crossover and mutation operators
In addition, it is important for adjusting the required network parameters in the design of dynamic systems. In order to decline trial-and-error, a self-tuning process is used the GA to compute significant parameters such as translation, dilation, weights, and membership functions. 2.5.2 Wavelet Network construction The wavelet neural networks are constructed by using the Multi Library Wavelet Neural Networks (MLWNN). This Library is constructed with many wavelet functions which are built by using the discretely dilated and translated version of the wavelet mother function<. In our approach the wavelet neural networks (WNN) are composed of three layers (an input layer, a hidden layer and an output layer) (Fig. 2). The hidden layer is composed with wavelet functions which are selected by using the Least Trimmed Square (LTS) method, which is used for initializing the wavelet neural networks. After the initializing of the WNN, the Genetic Algorithms (GA) is applied to train the WNN. The training is used to compute the dilation (ai), the translation (bi) and the weight (wi) values in order to optimize the WNN structure.
422
Abdesselem Dakhli et al. / Procedia Computer Science 96 (2016) 418 – 427
Figure 2 the three layer wavelet network
The wavelet neural network is constructed by using the Beta Wavelet function, which is defined by E(x)=Ex0,x1,p,q(x) [2526-27], x0 and x1 are real parameters.
E ( x, p, q, x0 , x1 )
°§ °¨ ®¨ °© ° ¯0
x x0 x c x0
· ¸ ¸ ¹
p
§ x x1 ¨ ¨ x x © 1 c
· ¸ ¸ ¹
q
if x >x0 , x1 @
otherwise
(12) We have proved, in [24-26-27], that all the derivatives of Beta function 䳲 L2() are of class C(Fig. 3). The general form of the nth derivative of Beta function is:
\ n ( x)
d n E ( x) d xn
ª « n n! p « 1 n 1 « x x 0 «¬
ª « n n n i ! p ¦ C in « 1 n 1 i « i 1 x x0 «¬
º » » E ( x ) P n ( x ) P1 ( x ) E ( x ) n 1» x x 1 »¼ º » n i !q » u P1 ( x ) E ( x ) n 1 i » x1 x »¼ n! q
(13) Where:
P
1
(x)
P x x0
x
1
q x (14)
P n ( x)
1
n
n! p
n! q
x x 0 n 1 x1 x n 1 (15)
If p = q, for all n N and 0 < n < p, the functions n(x) = dn(x)/dxn are wavelets [25-26].
423
Abdesselem Dakhli et al. / Procedia Computer Science 96 (2016) 418 – 427
Figure 3 First, second and third derivatives of Beta function
2.5.3 Proposed Learning Algorithm Our approach(Fig.4) uses GAs to build optimal WNN. It applies the GAs evolving to optimize the four WNN parameters (dilation parameter A, translation parameter B, Weight values from input nodes to hidden nodes V and Weight values from hidden nodes to output nodes W. The vector R(A,B,V,W) consists of 9 Weight values from input nodes to hidden nodes V= {vij, i= 1,2,. . ., N, where N= number of input nodes, j = 1,2, . . .,k, where k= number of hidden nodes}. 9 Weight values from hidden nodes to output nodes W = {wjz, j = 1,2, . . .,k and z = 1,2,. . .,m, where m = number of output nodes}. 9 Dilation parameters A = (a1,a2, . . .,dk). 9 Translation parameter B=(b1,b2,…,bk). The proposed Algorithm presents as follow (Fig.4): Step 1: The original data of DNA sequence is divided into training set and testing set for network training and testing, respectively. Step 2: Conversion of DNA sequence to genomic signal using a binary indicator and Power Spectrum Signal Processing Step 3: Construct a library W of discretely dilated and translated version of a given wavelet\ , according to the available training data set. x Apply the LTS methods to select optimal mother wavelet function (10)(11). x Choose, from the library, the N wavelet candidate that best matches an output vector. Step 3.1: Initializing initialize mother wavelet function library
<
^< i`, i
1 , 2 ,... m , m
<
, is the
number of wavelet function included, the LTS estimator is used to choose the best wavelet function and optimal wavelet function waveletBest= waveleti.. Step 3.2: Randomly initialize wjk and vij. Step 3.3: For k=1,…,m
a)
Calculate the predicted output y i via (3).
b) Compute the fitness (error) of each individual; the fitness function we are going to use is the so-called Normalized-Root-Mean-Squared-Error (NRMSE) : § · P y ¨¨ i yi ¸¸ / ¦ ( yi ¦ i 1© ¹ i1 P
NRMSEi
y)
2
.
(16)
c)
If the stopping criterion (Fitness (NRMSE) Threshold value ) is met, then stop; otherwise, go to the next step d) Find the ordered values NRMSEik d ... d NRMSEim
Step 4: Select the Best Tow Individual; the individual that produces the smallest error (NRMSE) has higher possibility to be selected Step 5: Apply Crossover Process on the Selected Individual P=(R1,…,RM) Ri=Ri min+rand(0,1)*(Ri max-Ri min) ni=Rc+F*(Ra-Rb) Go to the next step 3.3 opt
opt
ij
i
w ,a
Step 6: Create an empty matrix (feature_extraction)(
opt
and bi ) which has to contain the clusters of the DNA
sequences. Step 7: Let feature_extraction _DNA=
si ={
centers. Step 8: Randomly select ‘c’ cluster centers
opt
opt
opt
ij
i
i
w , a ,b
} be the set of data points and V={v1,v2,…,vc} be the set of
424
Abdesselem Dakhli et al. / Procedia Computer Science 96 (2016) 418 – 427
Step 9: Compute the distance between each data point and cluster centers. Step 10: Assign the data point to the cluster center whose distance from the group center is minimum of all the group centers. Step 11: Recalculate the new cluster center using: c
vi
§ 1 · i s ¨ c ¸¦ i © i¹j 1
(17)
Where, ‘ci’ represents the number of data points in ith cluster. Step 12: Recalculate the distance between each data point and new obtained cluster centers. Step 13: If no data point was reassigned then stop, otherwise repeat from step 10.
Figure 4 proposed approach
425
Abdesselem Dakhli et al. / Procedia Computer Science 96 (2016) 418 – 427
3. Results and Discussion The experiments were conducted on DNA sequences belonging to the Promoter gene, Primate splice-junction gene sequence, SPECTF-heart data and BUPA-liver disorders [28](Tab.1). Table 1. Datasets of DNA sequences Number of DNA sequences in the dataset 106
59
Primate splicejunction gene
3190
1595
SPECTFheart data
267
23
BUPA-liver disorders
345
7
Dataset Promoter gene
Average length of a DNA sequence in the dataset
To evaluate the performance of our approach, we have developed different experiments, each consisting of a different subset of test data. The classification comparative analysis is performed using a selection of published empirical datasets and synthetic DNA datasets [28]. Table 2. Distribution of Available Data into Training and Testing Set of DNA Sequence Dataset
Total
Training
Test
Promoter gene
106
60
46
Primate splicejunction gene
500
300
200
SPECTF-heart data
267
200
67
BUPA-liver disorders
345
200
145
Table 3. Analysis of all datasets with accuracy using ANN-back propagation[30] vs WNN-GA Accuracy using back propagation
Accuracy using GA(WNN-GA) (our method)
SNO.
Name of database
Training accuracy
Testing accuracy
Training accuracy
Testing accuracy
1
Promoter gene
85%
83%
98%
92%
2
Primate splice-junction gene
81%
85%
96%
94%
3
SPECTF-heart data
87%
84%
89%
93%
4
BUPA-liver disorders
77%
80%
99%
98%
426
Abdesselem Dakhli et al. / Procedia Computer Science 96 (2016) 418 – 427
Figure 5 Accuracy using ANN-back propagation vs WNN-GA
The results show that the WNN-GA can achieve very good prediction accuracy. The results of our approach tested on empirical datasets show that accuracy outperforms the other techniques in terms of percentage of the correct species identification. Table 3,4 and figure 5 show the distribution of the good classifications by class as well as the rate of global classification for all the DNA sequences of the validation phase. The results indicate the overwhelming supremacy of WNN-GA in accuracy as compared to ANN-back propagation. First, during the approximation phase, our model tried to decompose the input signal for every sequence and then tried to reconstruct the input signal. The estimation of the performance of this phase was measured by the Mean Square Error (MSE). Table 4 shows that the Mean Square Error (MSE) obtained are low (0.00132) and the run-time increases relatively with the size of the DNA sequence (2.632sec). The results show that the size of the DNA sequence increases the time of the training phase. Time depends on the size of the DNA sequence. When the size is equal to 1595, the training time is equal to 13.223 sec. Table 4 MSE of Approximation of the Signal for DNA DNA sequence
Seq. length
for each Class
MSE
Training
(Mean Square Error)
Time(sec)
Promoter gene
59
0.006689
10.666
Primate splicejunction gene
1595
0.00756
13.223
SPECTF-heart data
23
0.006852
6.965
BUPA-liver disorders
7
0.00132
2.632
The table 4 shows that the WNN-GA can achieve with good approximation of DNA sequences. 4. Conclusion In this paper, we used a new method of the training called Library Wavelet Neural Network Model. It is used to construct the Wavelet Neural Network. The WNN is used to approximate function f (x) of a DNA sequence signal. Our method depends on binary codification and Power Spectrum, processing the DNA sequence signal. Applying the k-means classification enables us to group the similar DNA sequences according to some criteria. This classification aims at distributing n DNA sequences characterized by p variables X1, X2…Xp in a number m of subgroups which are homogeneous as much as possible where every group is well differentiated from the others. In our approach, we used the Correlation Coefficient or Pearson Correlation Coefficient which is applied to measure the association between two vectors of DNA sequences signal. Our approach helps to classify organisms into different categories and classes which have significant biological knowledge and can justify the evolution and the identification of unknown organisms. The simulation results are demonstrated to validate the generalization ability and efficiency of the proposed WNN Model. These results were realized thanks to many capacities listed as; x The capacity of Library Wavelet Neural Network Model to construct the WNN. x The capacity of genetic algorithm and the LTS method to construct the optimal WNN.
Abdesselem Dakhli et al. / Procedia Computer Science 96 (2016) 418 – 427
x
It is concluded that besides being robust, WNN-GA is an effective method for solving the classification problems occurring in the classification of the DNA sequences..
Acknowledgements We would like to acknowledge the financial support, under the form of grant, from the General Direction of Scientific Research (DGRST), Tunisia. References [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28]
Q. ZHANG, A. Benveniste, Wavelet networks, IEEE Trans. on Neural Networks. 3(6), 1992, 889-898 J. Zhang, G. Walter, Y. Miao, et al. Wavelet neural networks for function learning, IEEE Trans. ON SIGNAL PROCESSING. 43(6), 1995, 1485-1497 Q. ZHANG, Using wavelet network in nonparametric estimation, IEEE Trans. on Signal Processing.8 (1997), 227-236 Y. C. Pati, P. S. Krishnaprasad, Analysis and synthesis of feed-forward neural networks using discrete affine wavelet transformations, IEEE Trans. on Neural Networks. 4 (1993), 73-85 S. A. Billings, H. L. Wei, A new class of wavelet networks for nonlinear system identification, IEEE Trans. on Neural Networks. 16 (2005), 862-874 J. H. Xu, D. W. C. Ho, A basis selection algorithm for wavelet neural networks, Neurocomputing, 48 (2002), 681-689 S. G. Mallat, Z. Zhifeng, Matching pursuits with time-frequency dictionaries, IEEE Trans. On Signal Processing. 41 (1993), 3397-3415 S. Chen, J. Wigger, Fast orthogonal least squares algorithm for efficient subset model selection, IEEE Trans. on Signal Processing. 43 (1995), 1513-1515 M. Han, J. Yin, The hidden neurons selection of the wavelet networks using support vector machines and ridge regression, Neurocomput. 72 (2008), 471479 S. G. Mallat, Z. Zhifeng, Matching pursuits with time-frequency dictionaries, IEEE Trans. On Signal Processing. 41 (1993), 3397-3415 Cathy H. Wu, Artificial neural networks for molecular sequence analysis, Compurers Chem. Vol. 21,(1997),No. 4. pp. 231-256. Agnieszka E Jach1 , Juan M Marín, Classification of Genomic Sequences via Wavelet Variance and a Self-Organizing Map with an Application to Mitochondrial DNA, Stat Appl Genet Mol Biol. Vol.9,(2010),pp.1544-6115.1562. Xiu Wen Yang, Jian Ping Li, Yuan Yan Tang, DNA Sequences Classification Based on Wavelet Packet Analysis, Lecture Notes in Computer Science Vol.2251,(2001), pp 424-429. Imen et al., Wavelet analysis of frequency chaos game signal: a time-frequency signature of the C. elegans DNA, EURASIP Journal on Bioinformatics and Systems Biology 2014, 2014:16 Wei D and Jiang Q. A DNA sequence distance measure approach for phylogenetic tree construction. Proceedings of the 2010 IEEE Fifth International Conference on Bio-Inspired Computing: Theories and Applications (BIC-TA). IEEE,(2010),pp. 204-212. Shi L and Huang H. DNA sequences analysis based on classifications of nucleotide bases. Springer Berlin Heidelberg ,(2012), 379-384. Bauer M, Schuster SM and Sayood K. The average mutual information profile as a genomic signature. BMC Bioinformatics 9,(2008), 48. Qi J, Wang B and Hao BI . Whole proteome prokaryote phylogeny without sequence alignment: a K-string composition approach. J. Mol. Evol.,(2004), 58: 1-11. J.P. Bao and R.Y. Yuan. A wavelet-based feature vector model for DNA clustering. Genetics and Molecular Research 14 (4),(2015), 19163-19152. S.bai Arniker and H.Keung Kawan, « Advanced Numerical representation of DNA sequences »,International Conference on Bioscience, Biochemistry and Bioinformatices IPCBEE, vol.31,(2012),pp.1-5. Randy L. Haupt, Sue Ellen Haupt, ”Practical Genetic Algorithms”, Willey- InterScience, Second Edition, chapter 2, 2004. Christian Blum, Jakob Puchinger, Gunther R. Raidl, Andrea Roli," Hybrid metaheuristics in combinatorial optimization: A survey", Applied Soft Computing 11,(2011), 4135–4151. Jan Roupec and Pavel Popela, The Nested Genetic Algorithms for Distributed Optimization Problems", Proceedings of the World Congress on Engineering and Computer Science 2011 Vol I, WCECS 2011. W. Bellil, C. Ben Amar and Mohamed AA: Synthesis of wavelet filters using wavelet neural networks”, Transactions on Engineering, Computation and Technology 2006, vol.13, (2006),pp. 108-111. C. Ben Amar, M. Zied and M. Adel Alimi, “Beta wavelets. Synthesis and application to lossy image compression”, Journal of Advances in Engineering Software, Elsevier Edition, vol.36, (2005), pp. 459 – 474. M. Zaied, O. Jemai and C. Ben Amar, “Training of the Beta wavelet networks by the frames theory: Application to face recognition”, ieeexplore.ieee.org, Image Processing Theory, Tools & Applications, (2008), 978-1-4244-3322-3/08/. C. Ben Amar, W. Bellil and M. Adel Alimi, “Beta Function and its Derivatives: A New Wavelet Family”, Transactions on Systems, Signals and Devices, vol.1, (2006), pp. 275-293. ftp site: genbank.bio.net
427
ScienceDirect Procedia Computer Science 96 (2016) 418 – 427
Wavelet Neural Networks for DNA Sequence Classification Using the Genetic Algorithms and the Least Trimmed Square Abdesselem Dakhlia*, Wajdi Bellilb, Chokri Ben amarc b
a Department of Computer Science, REGIM, University of Gabes 6002 Gabes, Tunisia Department of Electrical Engineering, REGIM, University of Gafsa 2110 Gafsa, Tunisia c Department of Electrical Engineering, REGIM, University Sfax 3018 Sfax, Tunisia
Abstract This paper presents a structure of the Wavelet Neural Networks used to classify the DNA sequences. The satisfying performance of the Wavelet Neural Networks (WNN) depends on an appropriate determination of the WNN structure optimization problem. In this paper we present a new method to solve this problem based on Genetic Algorithm (GA) and the Least Trimmed Square (LTS). The GA is used to solve the structure and the learning of the WNN and the LTS algorithm is applied to select the important wavelets. First, the scale of the WNN is managed by using the time-frequency locality of wavelet. Furthermore, this optimization problem can be solved efficiently by Genetic Algorithm as well as the LTS method to improve the robustness. The performance of the Wavelet Networks is investigated by detecting the simulating and the real signals in white noise. The main advantage of this method can guarantee the optimal structure of the WNN. The experimental results have indicated that the proposed method (WNN-GA) with the k-means algorithm is more precise than other methods. The proposed method has been able to optimize the wavelet neural network and classify the DNA sequences. Our goal is to construct a predictive approach that is highly accurate results. In fact, our approach allows avoiding the complex problem of form and structure in different groups of organisms. The experimental results are showed that the WNN-GA model outperformed the other models in terms of both the clustering results and the running time. In this study, we present our system which consists of three phases. The first one is the transformation, is composed of two sub steps; the binary codification of the DNA sequences and the Power Spectrum Signal Processing. The second step is the approximation; it is empowered by the use of the Multi Library Wavelet Neural Networks (MLWNN). Finally, the third one is the clustering of the DNA sequences, is realized by applying the algorithm of the k-means algorithm.
© 2016 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license © 2016 The Authors. Published by Elsevier B.V. (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under under responsibility Peer-review responsibilityofofKES KESInternational. International Keywords:Wavelet Neural Networks, GA, DNA sequences, LTS, MLWNN, K-means ;
1. Introduction The WNN have recently attracted extensive attention for its ability to identify effetely nonlinear dynamic systems with incomplete information [1-2-3-4-5]. The WNN were introduced by Zhang and Benveniste in 1992 as combination of the artificial neural network wavelet decomposition [1-2-3]. The generalization performance of the WNN is trained by least-square approach deteriorates when outliers are present. This training approach involves estimating parameters in the network by minimizing some function costs, a measure reflecting the approximation quality is performed by the network over the parameter space in the network. However, the studies of the WNN have often concentrated on small dimension [6]. The reason is that the complexity of network construction will be raised exponentially with the input dimension, i.e. the curse of dimensionality to improve the performance of the WNN in high dimension application. It is a key problem that how to appropriate determinate the network
* Corresponding author. Tel.: +21628002213; fax:+ 216 75 270 686 E-mail address: [email protected]
1877-0509 © 2016 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of KES International doi:10.1016/j.procs.2016.08.088
419
Abdesselem Dakhli et al. / Procedia Computer Science 96 (2016) 418 – 427
structure. This structure has been studied by several researchers. The research effort has been made to deal with this problem over the last decades [6-7-8-9]. The method, which is referred to as Matching Pursuit (MP), was first introduced by Mallat [10]. The Neural Network was used to solve the classification system, such as classification of the DNA sequences using the artificial neural networks [11]. Agnieszka E. et al used method to classify the genomic sequences. This method is combined a wavelet analysis and a self-organizing map algorithm [12]. It is used to extract feature of the oligonucleotide patterns of a sequence. The variation is quantified by the estimated wavelet variance, which yields a feature vector. This method is allowed the results to be visualized. When only thousands of nucleotides are available, wavelet-based feature vectors of short oligonucleotide patterns are more efficient in discrimination than frequency-based feature vectors of long patterns The Wavelet analysis is applied to extract the features of the DNA sequences [13]. The classification of two types of DNA sequence is studied. As well as, 20 samples of the artificial DNA sequences whose their types are known, are given in order to recognize the other types of the DNA sequences. The way of wavelet denoising keeps more main formation in the original curve but it cannot predict accurately the gas zones. In addition, the Wavelet analysis of frequency chaos game signal has been used to classify the DNA sequences. The results stemming from the complex Morlet wavelet analysis of the frequency chaos game signals have presented its accuracy in detection of variable DNA sequences structures. Moreover, this could serve in discovering unknown domains with potential biological significance in genomes [14-15-19]. This paper is organized as follow: in section II, we present an overview of the proposed method. In this section we present the transformation of the DNA sequences and the theory of wavelet neural networks. Section III deals with the simulation results of the proposed approach, which is used to classify the DNA sequences and Section IV ends up with a conclusion and a discussion. 2. Methods This paper presents a new method of clustering of the DNA sequences based on wavelet network using the Multi Library Wavelet Neural Networks (MLWNN) to approximate f(x) of the DNA sequence. The genetic algorithm is used to solve the structure and the learning of the WNN. This approach is divided into two stages: approximation of the input signal and classification of feature extraction of the DNA sequences using the Wavelet Neural Network (WNN) and the k-means algorithm. 2.1. Conversion of the DNA sequence into a genomic signal Our system is used the DNA sequences components to classify the species. These sequences are composed of four basic nucleotides (A, G, C and T) are called adenine, guanine, cytosine and thymine respectively, where each organism is identified by its DNA sequence [20-15-16]. The feature extraction of the DNA sequences can be viewed as finding a set of vectors which represents effectively information [17]. The method of the indicator translates the data into a digital format, which indicate the presence or absence of four nucleotides [18]. The binary indicator sequence is constructed to replace the individual nucleotides with values either 1 or 0. 0 stands for absence and 1 for presence of a particular nucleotide in specified location in DNA signal [19-20]. For example, if x[n] =[T T A T G TA ...], we obtain: x[n] =[0001 0001 1000 0001 0010 0001 1000 . ..] 2.2. Fourier Transform and Power Spectrum Signal Processing After the DNA sequence data have been constructed into these indicator sequences, they can be manipulated with mathematical function. The Fourier Transform is used to each indicator sequence x(n) and a new sequence of complex numbers, named f (x) , is obtained: N 1
¦
f (x)
n
X 0
e
(n )e
jS n / N
,k
0 ,1, 2 , ...N 1
(1)
It is easier to work with sequence Power Spectrum, rather than original discrete Fourier Transform. The Power Spectrum Se[k] is defined as, Se[ k ]
f ( x)
2
(2)
Where the frequencies k= 0, 1, 2, …, N-1. 2.3. Wavelet Neural Network and Time-frequency Analysis The combination of the wavelet transform and the artificial neuron networks defines the concept of the wavelet networks, which applied the mother wavelet functions instead of the traditional sigmoid function as a transfer function of the each neuron. It is composed of three layers, called the input layer, the hidden layer, and the output layer. It has the same structure as the architecture radial function. The salaries of the weighted outputs are added. Each neuron is joined to the other following layer.
420
Abdesselem Dakhli et al. / Procedia Computer Science 96 (2016) 418 – 427
The Wavelet network (Fig. 2) is defined by pondering a set of wavelets dilated and translated from one wavelet function with weight values to approximate a given signal f. Mapping functions to a time-frequency element training set, the overall response of the WNN is
N
^
¦
y
w
§ x bi · N i ¸ ¦ a © ai ¹ k 0
Z i< ¨
x
k
Where f is the regression function and the error term
H
i 1
k
(3) Where (x1, x2,…,xNi ) is the vector of the input, yˆ is the output of the wavelet neural networks, Nw is the number of wavelets, ai is the dilation, bi is the translation, and wi is the weight of the WNN. There, it is often useful to consider, the decomposition of wavelets cleanly, that the output can have a component refine in relation to the variables of coefficients ak (k = 0, 1... Ni) (Fig.2). wavelet occur in family of function and each is defined by translation bi which controls the position of a single function, called the mother wavelet <(x) and dilation ai which controls the scaling parameter. A WNN can be regarded as function approximator, which estimates an unknown function mapping:
f x H ,
y
(4) is a zero-mean random variable of disturbance.
2.4. Multi Library Wavelet Neural Network(MLWNN) The WNN is constructed by many approaches [2-3], Zhang applied two stages to construct the Wavelet neural Network: First, the discretely dilated and the translated version of wavelet mother function < are used to construct a wavelet library W:
®\ i :\ i ¯
W
x
½
D i\ ai x bi ,D i ¨ ¦ >\ ai xk bi @2 ¸ 2 , i 1,...L ¾, §
©k
·
n
1
¹
¿ (5) Where xk is the sampled input and L is the number of wavelets in W, then the best M wavelets is selected based on the training data from the library W, in order to construct the regression: 1
f x y ¦ w\ x , M
i
iI
i
(6)
Where I is a M-element subset wavelet from the wavelet library, and M d L. Secondly, the minimized of the cost function:
1 § · min ¦ ¨ y ¦ w\ x ¸ ¹ wi , iI n © n
j I
k
k 1
iI
i
i
k
2
(7)
,
Two heuristic algorithms are derived by Zhang, called, the stepwise selection by orthogonalization is used to select appropriate wavelet in the hidden units while the second one is used to choose the number of the hidden units and of the wavelets M, which are selected as the minimum of the so-called Akaike’s final prediction error criterion (FPE)[2-3]:
j
FPE
§ ¨ ¨ ©
f
· ¸ ¸ ¹
1
n
pa
1
n
pa
n
§ 1 n ¨ ¦¨ 2n k 1 ¨ f ©
x k y
· ¸2 ¸ k¸ ¹
n (8) Where npa is the number of the parameters in the estimator. The gradient algorithms are used to train the WNN, while the least mean squares are used to reduce the mean-squared error:
j w
· 1 n §¨ w ¸ 2 , ¦ ¸ n i 1 ¨© y i y ¹
(9) Where j is the real output from a trained WNN at the fixed weight vector w. In this study, the hidden layer is composed with wavelet functions which are selected by using the Least Trimmed Square (LTS) method, which is used for initializing the wavelet neural networks. The residual (or error) ei at the ith ouput of the WNN due to the ith example is defined by :
ei
y i yˆ i , i n
421
Abdesselem Dakhli et al. / Procedia Computer Science 96 (2016) 418 – 427
(10) The Least Trimmed Square estimator is used to select the WNN weights that minimize the total sum of trimmed squared errors :
E
total
1 2
p
l
¦ ¦ e k
1
i
2
(11)
ik
1
The notation for the ordered absolute residuals is e(i) , so
e 1 d e 2 d ... d e n .
2.5. Wavelet Network construction using Genetic Algorithm (GA) 2.5.1 Genetic Algorithm (GA) John Holland presented the Genetic Algorithm (GA) in 1975, is a computing search technique used in finding a solution in the optimization problems. GA applies the principles of evolution found in nature to the problem of finding an optimal solution [2122-23]. GA generates successive populations of the alternate solutions that are represented by a chromosome, i.e. a solution to the problem, until acceptable results are obtained. Each chromosome, individual solution, consists of number of binary code called genes. A fitness function assesses the quality of a solution in the evaluation step. The evolution from one generation to the next is performed using three operations: reproduction, crossover and mutation. The evaluation of the fitness value is used by the chromosomes for reproduction. The fitter chromosomes have higher priority to be chosen into the recombination pool according to the roulette wheel or the tournament selection methods. Crossover selects genes from two parent chromosomes using randomly chosen crossover point and creates two new off springs and mutation process changes chromosome randomly by altering a single bit as in Fig.1.
Figure 1 Crossover and mutation operators
In addition, it is important for adjusting the required network parameters in the design of dynamic systems. In order to decline trial-and-error, a self-tuning process is used the GA to compute significant parameters such as translation, dilation, weights, and membership functions. 2.5.2 Wavelet Network construction The wavelet neural networks are constructed by using the Multi Library Wavelet Neural Networks (MLWNN). This Library is constructed with many wavelet functions which are built by using the discretely dilated and translated version of the wavelet mother function<. In our approach the wavelet neural networks (WNN) are composed of three layers (an input layer, a hidden layer and an output layer) (Fig. 2). The hidden layer is composed with wavelet functions which are selected by using the Least Trimmed Square (LTS) method, which is used for initializing the wavelet neural networks. After the initializing of the WNN, the Genetic Algorithms (GA) is applied to train the WNN. The training is used to compute the dilation (ai), the translation (bi) and the weight (wi) values in order to optimize the WNN structure.
422
Abdesselem Dakhli et al. / Procedia Computer Science 96 (2016) 418 – 427
Figure 2 the three layer wavelet network
The wavelet neural network is constructed by using the Beta Wavelet function, which is defined by E(x)=Ex0,x1,p,q(x) [2526-27], x0 and x1 are real parameters.
E ( x, p, q, x0 , x1 )
°§ °¨ ®¨ °© ° ¯0
x x0 x c x0
· ¸ ¸ ¹
p
§ x x1 ¨ ¨ x x © 1 c
· ¸ ¸ ¹
q
if x >x0 , x1 @
otherwise
(12) We have proved, in [24-26-27], that all the derivatives of Beta function 䳲 L2() are of class C(Fig. 3). The general form of the nth derivative of Beta function is:
\ n ( x)
d n E ( x) d xn
ª « n n! p « 1 n 1 « x x 0 «¬
ª « n n n i ! p ¦ C in « 1 n 1 i « i 1 x x0 «¬
º » » E ( x ) P n ( x ) P1 ( x ) E ( x ) n 1» x x 1 »¼ º » n i !q » u P1 ( x ) E ( x ) n 1 i » x1 x »¼ n! q
(13) Where:
P
1
(x)
P x x0
x
1
q x (14)
P n ( x)
1
n
n! p
n! q
x x 0 n 1 x1 x n 1 (15)
If p = q, for all n N and 0 < n < p, the functions n(x) = dn(x)/dxn are wavelets [25-26].
423
Abdesselem Dakhli et al. / Procedia Computer Science 96 (2016) 418 – 427
Figure 3 First, second and third derivatives of Beta function
2.5.3 Proposed Learning Algorithm Our approach(Fig.4) uses GAs to build optimal WNN. It applies the GAs evolving to optimize the four WNN parameters (dilation parameter A, translation parameter B, Weight values from input nodes to hidden nodes V and Weight values from hidden nodes to output nodes W. The vector R(A,B,V,W) consists of 9 Weight values from input nodes to hidden nodes V= {vij, i= 1,2,. . ., N, where N= number of input nodes, j = 1,2, . . .,k, where k= number of hidden nodes}. 9 Weight values from hidden nodes to output nodes W = {wjz, j = 1,2, . . .,k and z = 1,2,. . .,m, where m = number of output nodes}. 9 Dilation parameters A = (a1,a2, . . .,dk). 9 Translation parameter B=(b1,b2,…,bk). The proposed Algorithm presents as follow (Fig.4): Step 1: The original data of DNA sequence is divided into training set and testing set for network training and testing, respectively. Step 2: Conversion of DNA sequence to genomic signal using a binary indicator and Power Spectrum Signal Processing Step 3: Construct a library W of discretely dilated and translated version of a given wavelet\ , according to the available training data set. x Apply the LTS methods to select optimal mother wavelet function (10)(11). x Choose, from the library, the N wavelet candidate that best matches an output vector. Step 3.1: Initializing initialize mother wavelet function library
<
^< i`, i
1 , 2 ,... m , m
<
, is the
number of wavelet function included, the LTS estimator is used to choose the best wavelet function and optimal wavelet function waveletBest= waveleti.. Step 3.2: Randomly initialize wjk and vij. Step 3.3: For k=1,…,m
a)
Calculate the predicted output y i via (3).
b) Compute the fitness (error) of each individual; the fitness function we are going to use is the so-called Normalized-Root-Mean-Squared-Error (NRMSE) : § · P y ¨¨ i yi ¸¸ / ¦ ( yi ¦ i 1© ¹ i1 P
NRMSEi
y)
2
.
(16)
c)
If the stopping criterion (Fitness (NRMSE) Threshold value ) is met, then stop; otherwise, go to the next step d) Find the ordered values NRMSEik d ... d NRMSEim
Step 4: Select the Best Tow Individual; the individual that produces the smallest error (NRMSE) has higher possibility to be selected Step 5: Apply Crossover Process on the Selected Individual P=(R1,…,RM) Ri=Ri min+rand(0,1)*(Ri max-Ri min) ni=Rc+F*(Ra-Rb) Go to the next step 3.3 opt
opt
ij
i
w ,a
Step 6: Create an empty matrix (feature_extraction)(
opt
and bi ) which has to contain the clusters of the DNA
sequences. Step 7: Let feature_extraction _DNA=
si ={
centers. Step 8: Randomly select ‘c’ cluster centers
opt
opt
opt
ij
i
i
w , a ,b
} be the set of data points and V={v1,v2,…,vc} be the set of
424
Abdesselem Dakhli et al. / Procedia Computer Science 96 (2016) 418 – 427
Step 9: Compute the distance between each data point and cluster centers. Step 10: Assign the data point to the cluster center whose distance from the group center is minimum of all the group centers. Step 11: Recalculate the new cluster center using: c
vi
§ 1 · i s ¨ c ¸¦ i © i¹j 1
(17)
Where, ‘ci’ represents the number of data points in ith cluster. Step 12: Recalculate the distance between each data point and new obtained cluster centers. Step 13: If no data point was reassigned then stop, otherwise repeat from step 10.
Figure 4 proposed approach
425
Abdesselem Dakhli et al. / Procedia Computer Science 96 (2016) 418 – 427
3. Results and Discussion The experiments were conducted on DNA sequences belonging to the Promoter gene, Primate splice-junction gene sequence, SPECTF-heart data and BUPA-liver disorders [28](Tab.1). Table 1. Datasets of DNA sequences Number of DNA sequences in the dataset 106
59
Primate splicejunction gene
3190
1595
SPECTFheart data
267
23
BUPA-liver disorders
345
7
Dataset Promoter gene
Average length of a DNA sequence in the dataset
To evaluate the performance of our approach, we have developed different experiments, each consisting of a different subset of test data. The classification comparative analysis is performed using a selection of published empirical datasets and synthetic DNA datasets [28]. Table 2. Distribution of Available Data into Training and Testing Set of DNA Sequence Dataset
Total
Training
Test
Promoter gene
106
60
46
Primate splicejunction gene
500
300
200
SPECTF-heart data
267
200
67
BUPA-liver disorders
345
200
145
Table 3. Analysis of all datasets with accuracy using ANN-back propagation[30] vs WNN-GA Accuracy using back propagation
Accuracy using GA(WNN-GA) (our method)
SNO.
Name of database
Training accuracy
Testing accuracy
Training accuracy
Testing accuracy
1
Promoter gene
85%
83%
98%
92%
2
Primate splice-junction gene
81%
85%
96%
94%
3
SPECTF-heart data
87%
84%
89%
93%
4
BUPA-liver disorders
77%
80%
99%
98%
426
Abdesselem Dakhli et al. / Procedia Computer Science 96 (2016) 418 – 427
Figure 5 Accuracy using ANN-back propagation vs WNN-GA
The results show that the WNN-GA can achieve very good prediction accuracy. The results of our approach tested on empirical datasets show that accuracy outperforms the other techniques in terms of percentage of the correct species identification. Table 3,4 and figure 5 show the distribution of the good classifications by class as well as the rate of global classification for all the DNA sequences of the validation phase. The results indicate the overwhelming supremacy of WNN-GA in accuracy as compared to ANN-back propagation. First, during the approximation phase, our model tried to decompose the input signal for every sequence and then tried to reconstruct the input signal. The estimation of the performance of this phase was measured by the Mean Square Error (MSE). Table 4 shows that the Mean Square Error (MSE) obtained are low (0.00132) and the run-time increases relatively with the size of the DNA sequence (2.632sec). The results show that the size of the DNA sequence increases the time of the training phase. Time depends on the size of the DNA sequence. When the size is equal to 1595, the training time is equal to 13.223 sec. Table 4 MSE of Approximation of the Signal for DNA DNA sequence
Seq. length
for each Class
MSE
Training
(Mean Square Error)
Time(sec)
Promoter gene
59
0.006689
10.666
Primate splicejunction gene
1595
0.00756
13.223
SPECTF-heart data
23
0.006852
6.965
BUPA-liver disorders
7
0.00132
2.632
The table 4 shows that the WNN-GA can achieve with good approximation of DNA sequences. 4. Conclusion In this paper, we used a new method of the training called Library Wavelet Neural Network Model. It is used to construct the Wavelet Neural Network. The WNN is used to approximate function f (x) of a DNA sequence signal. Our method depends on binary codification and Power Spectrum, processing the DNA sequence signal. Applying the k-means classification enables us to group the similar DNA sequences according to some criteria. This classification aims at distributing n DNA sequences characterized by p variables X1, X2…Xp in a number m of subgroups which are homogeneous as much as possible where every group is well differentiated from the others. In our approach, we used the Correlation Coefficient or Pearson Correlation Coefficient which is applied to measure the association between two vectors of DNA sequences signal. Our approach helps to classify organisms into different categories and classes which have significant biological knowledge and can justify the evolution and the identification of unknown organisms. The simulation results are demonstrated to validate the generalization ability and efficiency of the proposed WNN Model. These results were realized thanks to many capacities listed as; x The capacity of Library Wavelet Neural Network Model to construct the WNN. x The capacity of genetic algorithm and the LTS method to construct the optimal WNN.
Abdesselem Dakhli et al. / Procedia Computer Science 96 (2016) 418 – 427
x
It is concluded that besides being robust, WNN-GA is an effective method for solving the classification problems occurring in the classification of the DNA sequences..
Acknowledgements We would like to acknowledge the financial support, under the form of grant, from the General Direction of Scientific Research (DGRST), Tunisia. References [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28]
Q. ZHANG, A. Benveniste, Wavelet networks, IEEE Trans. on Neural Networks. 3(6), 1992, 889-898 J. Zhang, G. Walter, Y. Miao, et al. Wavelet neural networks for function learning, IEEE Trans. ON SIGNAL PROCESSING. 43(6), 1995, 1485-1497 Q. ZHANG, Using wavelet network in nonparametric estimation, IEEE Trans. on Signal Processing.8 (1997), 227-236 Y. C. Pati, P. S. Krishnaprasad, Analysis and synthesis of feed-forward neural networks using discrete affine wavelet transformations, IEEE Trans. on Neural Networks. 4 (1993), 73-85 S. A. Billings, H. L. Wei, A new class of wavelet networks for nonlinear system identification, IEEE Trans. on Neural Networks. 16 (2005), 862-874 J. H. Xu, D. W. C. Ho, A basis selection algorithm for wavelet neural networks, Neurocomputing, 48 (2002), 681-689 S. G. Mallat, Z. Zhifeng, Matching pursuits with time-frequency dictionaries, IEEE Trans. On Signal Processing. 41 (1993), 3397-3415 S. Chen, J. Wigger, Fast orthogonal least squares algorithm for efficient subset model selection, IEEE Trans. on Signal Processing. 43 (1995), 1513-1515 M. Han, J. Yin, The hidden neurons selection of the wavelet networks using support vector machines and ridge regression, Neurocomput. 72 (2008), 471479 S. G. Mallat, Z. Zhifeng, Matching pursuits with time-frequency dictionaries, IEEE Trans. On Signal Processing. 41 (1993), 3397-3415 Cathy H. Wu, Artificial neural networks for molecular sequence analysis, Compurers Chem. Vol. 21,(1997),No. 4. pp. 231-256. Agnieszka E Jach1 , Juan M Marín, Classification of Genomic Sequences via Wavelet Variance and a Self-Organizing Map with an Application to Mitochondrial DNA, Stat Appl Genet Mol Biol. Vol.9,(2010),pp.1544-6115.1562. Xiu Wen Yang, Jian Ping Li, Yuan Yan Tang, DNA Sequences Classification Based on Wavelet Packet Analysis, Lecture Notes in Computer Science Vol.2251,(2001), pp 424-429. Imen et al., Wavelet analysis of frequency chaos game signal: a time-frequency signature of the C. elegans DNA, EURASIP Journal on Bioinformatics and Systems Biology 2014, 2014:16 Wei D and Jiang Q. A DNA sequence distance measure approach for phylogenetic tree construction. Proceedings of the 2010 IEEE Fifth International Conference on Bio-Inspired Computing: Theories and Applications (BIC-TA). IEEE,(2010),pp. 204-212. Shi L and Huang H. DNA sequences analysis based on classifications of nucleotide bases. Springer Berlin Heidelberg ,(2012), 379-384. Bauer M, Schuster SM and Sayood K. The average mutual information profile as a genomic signature. BMC Bioinformatics 9,(2008), 48. Qi J, Wang B and Hao BI . Whole proteome prokaryote phylogeny without sequence alignment: a K-string composition approach. J. Mol. Evol.,(2004), 58: 1-11. J.P. Bao and R.Y. Yuan. A wavelet-based feature vector model for DNA clustering. Genetics and Molecular Research 14 (4),(2015), 19163-19152. S.bai Arniker and H.Keung Kawan, « Advanced Numerical representation of DNA sequences »,International Conference on Bioscience, Biochemistry and Bioinformatices IPCBEE, vol.31,(2012),pp.1-5. Randy L. Haupt, Sue Ellen Haupt, ”Practical Genetic Algorithms”, Willey- InterScience, Second Edition, chapter 2, 2004. Christian Blum, Jakob Puchinger, Gunther R. Raidl, Andrea Roli," Hybrid metaheuristics in combinatorial optimization: A survey", Applied Soft Computing 11,(2011), 4135–4151. Jan Roupec and Pavel Popela, The Nested Genetic Algorithms for Distributed Optimization Problems", Proceedings of the World Congress on Engineering and Computer Science 2011 Vol I, WCECS 2011. W. Bellil, C. Ben Amar and Mohamed AA: Synthesis of wavelet filters using wavelet neural networks”, Transactions on Engineering, Computation and Technology 2006, vol.13, (2006),pp. 108-111. C. Ben Amar, M. Zied and M. Adel Alimi, “Beta wavelets. Synthesis and application to lossy image compression”, Journal of Advances in Engineering Software, Elsevier Edition, vol.36, (2005), pp. 459 – 474. M. Zaied, O. Jemai and C. Ben Amar, “Training of the Beta wavelet networks by the frames theory: Application to face recognition”, ieeexplore.ieee.org, Image Processing Theory, Tools & Applications, (2008), 978-1-4244-3322-3/08/. C. Ben Amar, W. Bellil and M. Adel Alimi, “Beta Function and its Derivatives: A New Wavelet Family”, Transactions on Systems, Signals and Devices, vol.1, (2006), pp. 275-293. ftp site: genbank.bio.net
427