CONTENTS:I. ABSTRACT II. INTRODUCTION III. EARLIER MEASURES IV.

NEWER APPROACHES a) b) c)
USING NEURAL NETWORKS USING CELLULAR AUTOMATA DISTRIBUTED DATABASES

2 3 5 7
7 13 18

V. TWO STAGE CLASSIFIER VI. FUTURE WORKS VII. CONCLUSION VIII. REFERENCES
1|Pa ge

21 23 24 24

PROTEIN CODING REGION IDENTIFICATION
Abstract:- Genes carry the instructions for making proteins that are found in a cell as a specific sequence of nucleotides in DNA molecules. These specific sequences of bases encode instructions on how to make protein. But the regions of gene that code for proteins may occupy only a small region of a sequence. The problem of gene recognition is to define an algorithm which takes as input DNA sequence and produces as output a feature table describing the location and structure of the patterns making up any genes present in the sequence. Identifying the coding region is of vital importance in understanding these genes. Here I have given certain pattern classifiers which will the test data set presented by fickett and tung to identify protein coding regions. Firstly we have considered some of the test data set using earlier measures for PCRI. We have described some of them not all Then we have proposed both Feed Forward connectivity and Backpropagation approach of neural networks for test data identification. We have found a highly accurate but space consuming CRM module. We have compared its results Then we propose cellular automata based pattern classifier to identify the coding region of a DNA sequence both in centralized and distributed databases. CA is simple efficient and produces more accurate classifier than that have previously been obtained for a range of different sequence lengths. And we have compared some of the older measures with FMACA based pattern classifier. At the end a genetic algorithm based two-stage pattern classifier was used to solve the PCRI. And although implementation was not done but we have found its effectiveness in various other data bases and suitable work can be done in this direction.

2|Pa ge

A protein coding gene basically found in a t -RNA It is any pattern in a DNA sequence It results in the generation of protein product through a process known as transcription. Our objective is to define an algorithm which takes an input DNA sequence and produces an output a feature table describing the location and structure. Fig 1:.Protein Synthesis in a DNA 3|Page . There are two types of regions Coding and Non-coding.N C N h t is P otein Coding egion dentifi tion? It is a technique to identify the coding (EXONS) and non coding (INTRONS) region in a protein coding base.

. 4. How to build an algorithm for PCRI? A natural overall approach to building gene recognition algorithms is to first construct component algorithms that recogni e the major features of genes: statistical bias in Exon sequence. 2. 3. It i i .How to Develop an Algorithm 4|Page . promoters. enhancers.bje tives of PC :1. Fig 2:. etc. and then to build a combined algorithm that recogni es when all these component patterns occur in a pattern consistent with that present in a gene . but for a systematic approach to building a comprehensive recognition algorithm one thing is still missing. Finding Protein coding regions these sequences biochemicall is time consuming and expensi e. It can be a good mechani m for genetic engineering. namely an objective comparison and evaluation of the competing recognition techniques. A great many ideas have been suggested for recognition of the components of genes. Possibilit exists of using computational procedures to anal l i i i t t i t i e raw base sequence information and to predict coding regions (as well as other signal regions) by the use of duplicate fragments. the patterns at Intron junctions.

T}.The extent to which the nucleotide is distributed over the three codon positions.C. Dicodon Measure:.They are offset by 1& 2 and similar to dicodon measure III. Run Measure: .(b) )2 VI.G. i ) . Position asymmetry measure: .i) is the frequency of base b over position I and (b) = ™i f( b.ARL R MEASURES:A.The Run feature is a vector of length 14 for each non trivial subset S = {A. Hyper plane Equation was used. Hexamer-1 & 2 :. IV. II.It is simply the longest sequence of codons that do not contain a stop codon. Diamino Acid Usage measure :. Codon usage Measure :. Measures for Protein Coding regions Identifications Definition:. VII.A coding measure is a vector of measurements on a DNA sequence.It is again a vector of the 440 amino acids frequency which are obtained by translating from nucleotide sequence to an amino acid string Fig 3:. i )/3 Asymmetry is defined as assymm(b) = ™i ( f( b . It contains an entry which gives longest contiguous subsequence that has all entries from s V.Training Set of Fickett And Tung DataBase 5|Page . If f(b.The codon usage feature is a vector of 64 codon frequencies. According to Fickett and tung there are 21 coding measures I. Open Reading Frame Measure:.It is the vector of 4096 dicodon frequencies of every possible dicodon where the dicodon counts are accumulated at sights whose starting point is multiple of 3.

B. Types of approaches Direct approach: Look for stretches that can be interpreted as protein using the genetic code Example: Open reading Frame measure Statistical approaches: Use other knowledge about likely coding regions Example: Dicodon usage Fig 4:-Open Reading Frame Measure Calculation Windows: Many sequence analyses require calculating some statistic over a long sequence looking for regions where the statistic is unusually high or low To do this. we define a window si e to be the width of the region over which each calculation is to be done 6|Page .

NEWER APPROACHES:1. II. they consist of many units connected to each other by variable strength links layer and one µoutput¶ layer ± and are known as µperceptrons¶. Learning:. VII. 7|Pa ge . V. Archi ecture :. So What we get a measure for learning. VI.The errors incurred during learning are propagated back through the network. Once trained. Unsupervised learning:. Using neural networks:I. Defi i i :.It is a number associated with a node which decreases or increases with each learning. Processing:-The unit (also known as a neuron or node) is the main processing element of the neural network. This training is known as supervised learning IV. the neural network should be capable of predicting an output given a previously unseen set of inputs. Back Propagation:.Name of this network is due to the direction of flow of the data. VIII.the task for the network is to relate the variables it receives in the input layer to some desired behaviour at the output layer by repeatedly presenting the examples to the network in a process known as training.The simplest neural networks there are only two layers ± one µinput¶ III.Neural networks were originally conceived as computational models of the way in which the human brain works. Weights:.Unsupervised learning does not require a desired response (and therefore output data) to train and is concerned with discovering a common set of features across the input data in one class or pattern. combined with an internal function. converts the input signals to an output signal. It receives a set of input signals and. Feed-Forward Approach:. Like the human brain. But we can have multiple layer also.

i3 . i ) and the desired output (o) data to the network 3.e. . i2. 2. Calculate the output from the network using the expression: f (s) 4.PERCEPTRON LE RNING RULE 1. (c) if output is 1 and should be 0 w (t + 1) = w (t) . . Adapt the weights: (a) if correct w (t + 1) = w (t). after updating) and f (s) is the function used to compute the output in the network 8|Page . ¡ ¡ ¡ ¡ ¡ ¡   ¡ ¡ Fig 1:. (b) if output is 0 and should be 1 w (t + 1) = w (t) + i (t). Present the input data (1. Initiali e the network weights and unit thresholds to some small random values.i (t).Preceptron Learning Rule Where w (t) is the weight value at time t. w (t + 1) is the weight value at time t + 1 (i.

3. The output from the final layer is the vector of output values. 4.Backpropagation:1. starting from the output layer and moving backwards using the equation Wij (t + 1) = wij (t) + pjopj d is the desired output. Adapt the weights. . . i ) and the desired output (o) data to the network. The sum shown for hidden layers is over the k nodes in the layer above the current layer for which the pj will already have been computed Fig 6:.BackPropagation 9|Page ¢ . Calculate the output from the network using the expression f (s) ] for each layer in the network. i2. i3 . Present the input data (i1. 2. o is the actual output c is a constant used in the sigmoid function. Initiali e the thresholds and weights.

Feed Forward Connectivity:I.s sixsi Where si represents the probability of codon s at position i p (T {Xi})=1/(1 + exp(™sTsNs + )) Ts = log ( qs + ps ) And = log (p (f)/p (t)) Fig 7:.Feed Forward Connectivity Fig 1:.Protein Synthesis in a DNA Fig 1:. III. Assuming that individual codon positions are independent of the codons p({Xi} T)=šI. Given a collection of examples labeled true or false for exons it is true for introns it is false II. We can employ a supervised Neural Network by using Bayes theorem p(T {Xi}) = p({Xi} T)p(T)/{p({Xi} T)p(T)+p({Xi} F)p(F)} Here P(t) and p(f) are the priori probabilities of the true class and false class.Protein Synthesis in a DNA 10 | P a g e .

CRM Based Approach:1. CRM means Coding Recognition Module.0 are then evaluated. To determine the likely hood of a given sequence position a program written in C language is used which calculates the sensors Values for a 99 base sequence and the sensor signals between 0. The Coding Module is used to find exons or parts of exons that encode protein.0 and 1. 5. In the following table we will see that it is a very accurate algorithm Fig 10:. 6. and an output node. 2. 2 Hidden layers of 14 and 5 nodes . This network is consisting of 7 input nodes .Using A CRM Module 11 | P a g e . 3. A back propagation neural net was created to integrate the output from the 7 sensors Algorithms. 4.

Table showing a Database comparison for CRM NOTE:. 12 | P a g e .Fig 11:.we see that our classifier almost gives a 99% accuracy.

‡ The transient states finally settles down to one of the attractor basins after passing through a lot of intermediate states. FMACA:‡ A FMACA is a special class of FCA that can efficiently model an associative memory to perform pattern recognition/classification task ‡ Its state transition behaviour consists of multiple components . qi+1(t) ) Fuzzy Cellular Automata:An elementary FCA is a linear array of cells which evolves in time. the next state qi(t+1) of a cell is assumed to be dependent only on itself and on its two neighbours (left and right). ‡ Each component. The global evolution results from synchronous application of the local rule to all cells of the array. and is denoted as qi(t+1)=f(qi-1(t) . Fig 12:. as noted in is an inverted tree. ‡ A cycle in a component is referred to as an attractor.1] (fuzzy states) and changes its state according to a local evolution function on its own state and states of its two neighbours. qi(t) . Each cell of the array assumes a state qi. each rooted on a cyclic state.2. Using Cellular Automata:A Cellular Automaton (CA) can be viewed as an autonomous finite state machine (FSM) consisting of a number of cells. In a 3-neighborhood dependency.State diagram of 3 cell 3 state FMACA 13 | P a g e . a rational value in the interval [0.

FMACA BASED K-means clustering Algorithm:Algorithm 1:Step 1. We show the algorithmic approach on building a tree classifier. Step 3. Stop. Otherwise if S/ belongs to more than one class. The elements of the training set S get distributed into k attractor basins/nodes.e. Let S/ be the set of elements in an attractor basin. 14 | P a g e . Step 4. Step 5. all patterns of S/ are covered by an attractor basin/node belonging to only one particular class). then mark it as a leaf node and label that attractor basin/node as that class. then repeat the steps 1 to 4. Step 2. If S/ belongs to any one particular class (i. FMACA with k attractor basins is generated.

Errors and other Benchmarks:The objective function (which is to be minimized) is the sums of squares distances of each DNA sequence and its assigned cluster centres.Implementation:For using a FMACA based classifier we need a limited window for eg 100 base pairs which can be used for standard experimental observations and it computes the features of that window alone. Low error rate leading to maximum classification accuracy. An important issue in applying FMACA works to DNA sequence classification is how to encode DNA sequences as the input of the FMACA based tree structure classifier. T}. From a one dimensional point of view a DNA sequence contain characters from a 4-letter base pair S = {A. C. The goal is to identify whether the entirely all-coding or all-non coding. Thus good implementation must be the focus. SSdistances =™ x-C(x)2] where C(x) is the mean of the cluster that DNA position x is assigned to Minimizing the SSdistances is equivalent to minimizing the Mean Squared Error (MSE) Wrong to state that lesser MSE means better classifications The basic criteria for FMACA tree design are: 1. We have followed this procedure in our study and we have used the benchmark human dna database developed fickett and tung. Small tree height leading to low retrieval time. Less number of nodes leading to low memory overhead 3. 2. Time complexity is O(n) By using Dependency Vector scheme 15 | P a g e . G.

And also takes very less memory overhead and retrieval time.Here we have compared with some of the other measures and found out that our FMACA based classifier is indeed a very accurate classifier. We have shown that in the graphs below. 16 | P a g e .

Fig 12:.FMACA INTERFACE 17 | P a g e .

` the base classifier at each local site must have high classification accuracy ` it must be scalable to handle disk resident datasets. ` It should have high throughput with less storage requirements ` Volume of data transferred from each local to central site should be minimum ` Aggregation of base classifiers must induce globally meaningful knowledge.DISTRIBUTED DATABASES ` Distributed classifiers work in ensemble learning approach ` Multiple classifiers as base classifiers at different data sources. 18 | P a g e . ` Results derived out of base classifiers are next combined at central location to enhance the classification accuracy of predictive models. ISSUES IN DISTRIBUTED DATABASES ` The aggregation of the base classifiers ` The communication cost from different local data sources to the central location where final classification is performed.

We have shown here algorithmic approach for distributed database PCRI 19 | P a g e .

For Distributed Datasets: Total training set is equally distributed to each local site and then same scheme is employed at each local site to synthesize the FMACA.min. The dataset also includes the reverse complement of every sequence. the possible value range of an attribute attri is (attrvali. then the normalized value of attrvalij i £ 20 | P a g e . Coding Measure: Typically. non-overlapping human DNA sequences of length 54 have been extracted from all human sequences. Suppose. a coding measure is a vector of measurements on a DNA subsequence. The tree then classifies the Test Set. For the first dataset. and the real value that class element j takes at attri is attrvalij. with shorter pieces at the ends discarded. Every sequence is labeled according to whether it is entirely coding. 2. The FMACA based base classifiers (BCi s) are next collected in the centralized location where predictive accuracy is tested with the test set. There are 21 coding measures. or mixed. entirely non-coding. and the mixed sequences are discarded. attrvalj. 3. which makes the problem of identifying coding regions somewhat harder. Comparison: All the attributes in a dataset are normalized to facilitate FMACA learning. The benchmark human data include three different datasets. The predictive accuracy of the proposed scheme is evaluated through the majority voting scheme. They have surveyed and compared it.max). 4.Experimental observation 1. For Centralized Datasets: There is use of a training set exclusively to construct a FMACA tree based structured pattern classifier. This means that one-half of the data is guaranteed to be from the non-sense strand of the DNA.

21 | P a g e .

We see as in below GA employed will significantly decrease the memory overhead 22 | P a g e .

Use of GA in FMACA classifiers 3. Can be used to distinguish between different kinds of genes.FUTURE WORKS 1. Use of SVM {Support Vector Machines}. Use of autoregressive modelling approach for introducing better Protein coding region identification. Can be used to genetically build protein Which can ensure higher immunity 3.CANCER Treatment by Knowing The protein Code 23 | P a g e . 2. Use of NNt Tree classification For MLP(multi layer perceptron) for better 4. ADVANTAGES 1. Can be used successfully to locate a cancerous protein coding region 2. Fig 13:.

1982. Acad." Proc. Flocchini. vol.CONCLUSION:So hereby conclude that by using efficient PCRI we can successfully make corrections to faulty genes. Sikdar and P. 2. "Recognition of protein coding regions in dna sequences. Biplab K. E88-D. 9. A study of fuzzy and many-valued logics in cellular automata. R. 691-702. ³Basins of Attraction of Cellular Automata Based Associative Memory and its Rule Space´ Pradipta Maji.Pal Chaudhuri. IEICE Volume 4. 88. Sci. in Proc. Santoro. vol. 2000. and N. "Locating protein-coding regions in human dna sequences by a multiple sensor-neural network approach." IEICE Transactions on Information and Systems. pp. "Assessment of protein coding measures. pp. Natl. ³Fuzzy Cellular Automata Based Associative Memory for Pattern Recognition´ Pradipta Maji and Chandra Das .Pal Chaudhuri. 1285-1289. Angelo B. 12. Heather Betel and Paola Flocchini. 11.Pal Chaudhuri. 1126111265. Tung. Maji and P. Chandrama Shaw. G. vol. REFERENCES:1. vol. Mol. 1991 P. 6441-6450. J." J. P Pal Choudhary ³Cellular Automata in protein coding Region Identification ³ 9proceedings of ICISIP-2005 24 | P a g e . F. S. It is not far when our dream of becoming immortal and space travellers will come true. Chaudhuri. Fickett. Uberbacher and R. "Fuzzy Cellular Automata For Modeling Pattern Classifier. 4. 5. Farber. Sumanta Das. ³Fuzzy Cellular Automata for Modeling Pattern Classifier´. "Convergence and Aperiodicity in Fuzzy Cellular Automata: Revisiting Rule 90. Fuzzy cellular automata and their chaotic behavior. International Symposium on Nonlinear Theory and its Applications. Knowledge Discovery in Distributed Biological Datasets Using Fuzzy Cellular Automata Pradipta Maji.. Hawaii... and K." Physica D. Flocchini. 10. 1992. Thereby a lot of new opportunities emerge. vol. Il. Mauri. 7. Cattaneo. and N. Geurts. Mural. Wolfram Media. pp. E. Santoro (1993). J." Nucleic Acids Res. P. 471-479. Mingarelli. Mingarelli(2007). S. Lapedes. P.. P. 6. April 2005. A. 226. Niloy Ganguly. Champaign. 16. 13. Sirotkin. Biol. A New Kind of Science (2002). A. 20. Wolfram. USA. 1992. G.Pal Chaudhuri. Fickett and C. 14. 10. pp. Pradipta Maji and P. 5303-5318. 15." Nucleic Acids Res. ³ ³Theory and Application of Cellular Automata for Pattern Classification´ Pradipta Maji and P. "Determination of eucaryotic protein coding regions using neural networks and information theory. Samik Barua. 8. ´On The Relationship Between Boolean and Fuzzy Cellular Automata´ Pradipta Maji and P. 3. pp.

25 | P a g e .

Sign up to vote on this title
UsefulNot useful