The sequence alignment of three or more biological sequences such as the Protein, DNA or RNA (Auyeung and Melcher, 2005[3]; Wei et al., 2013[66]) is known as the multiple sequence alignment (Hamidi et al., 2013[21]). One of the standard techniques in bioinformatics for reviling the relationship between collections of evolutionarily or structurally related protein is sequence alignment.

Sequence alignment are extensively be used for improving the secondary and tertiary structure of protein and RNA sequences, which is used for drug designing and also to find distance between organism. In MSA, the foremost effort is made to find the optimal alignment for a group of biological sequences. In the past research, we have observed several reliable and efficient techniques for alignment of multiple sequences, which includes evolutionary algorithm (GA) (Peng et al., 2011[51]), HMM (Eddy, 1998[13]) and the generic probabilistic metaheuristic for the global optimization problem (Kirkpatrick et al., 1983[32]).

One of the widely studied branches in bioinformatics is sequence similarity, also known as a subset of sequence analysis. The available molecular sequence data have enough resources that can teach us about the structure, function and evolution of biological macromolecules. The main objective of an MSA is to align sequences which can show the biological relationship between the input sequences, but to develop a reliable MSA program is never easy. In general the MSA problem can be seen as: Let N number of sequences is supplied as input with a predetermined scoring scheme for finding the best matches among the letters (as every sequences consists of a series of letter). Although, definition stated here is simple but still it requires certain input such as the selection of input sequence and comparison model along with the optimization of the model to get completed in all respect. There are various issues demonstrated in the literature (Aniba et al., 2010[1]; Pop and Salzberg, 2008[53]; Sellers,1984[55]) for alignment of protein sequences. First, the protein family described in sequences databases have complex multi domain architecture with huge unstructured regions. Second, the new sequences selected through automatic methods contains relevant amount of sequence error (Yonghua et al., 2004[70]; Wen and Tan, 1996[68]).

There are various methods which can be used to solve MSA problem such as the iterative (Mohsen et al., 2007[39]) classical, progressive algorithms (Kupis and Mandziuk, 2007[33]). All these algorithms are based on global or local alignment (Wei et al., 2013[66]; Changjin and Tewfik, 2009[8], Ankit and Huang, 2008[2]) techniques. The Global alignment technique, aids in making the sequences aligned from end to end points. Whereas, the local alignment technique first identifies a substring within a string and then tries to align it with the target string.

In general, local alignment is considered for sequence alignment but some time it creates problem because here in local alignment we have to deal with an additional challenge of identifying the regions of similarity. A dynamic programming based approach which are mostly used as the local and global alignment technique is the Smith-Waterman algorithm (Haoyue et al., 2009[22]) and Needleman-Wunsch algorithm (Needleman and Wunsch, 1970[43]). The dynamic programming (DP) (Zhimin and Zhong, 2013[72]) approach are considered to be good alignment option for not more than two sequences. Here, one thing is to be noted that MSA is a combinatorial problem (NP-hard) (Kececioglu and Starrett, 2004[30]) and when the number of sequences increases the computational effort becomes prohibitive. Feng and Doolittle (1987[15]) proposed a progressive alignment algorithm (tree-base algorithm), which uses the method of Needleman and Wunsch and for constructing an evolutionary tree (Bhattacharjee et al., 2006[4]) to know the relationship between sequences. The progressive alignment algorithms perform it operation through branching order of a guide tree and thus often get trapped to local optima (Naznin et al., 2012[42]). To avoid such kind of local optima it is suggested in the literatures to use either stochastic or iterative procedure (Mohsen et al., 2007[39]; Gotoh, 1982[19]).

By referring to various literature studies (Devereux et al., 1984[11]; Jagadamba et al., 2011[27]; Nguyen and Yi, 2011[45]; Katoh et al., 2005[29]; Pei and Grishin, 2007[50]; Li et al., 2004[35], Ma et al., 2002[37]; Pearson, 2000[49]), it can be concluded that none of the existing algorithms were accurate enough to provide an optimal alignment for all the datasets. As a result, with the uses of iterative refinement strategies (Gotoh, 1982[19]), Hidden Markov Models (Eddy, 1998[13]) or Genetic Algorithms (Peng et al., 2011[51]) an iterative algorithms (Mohsen et al., 2007[39]) were developed to construct more reliable and efficient multiple alignments. Also, all these methods listed above have shown their superiority in aligning distantly related sequences for a variety of datasets (Blackshields et al., 2006[5]; Thompson et al., 1999[62]). However, some accuracy was degraded while considering the distantly related sequences.

The above paragraph gives a clear indication that none of the method listed above can provide an accurate or meaningful alignment in all possible situations, irrespective of their advantages or disadvantages. Progressive alignment methods are known to be very fast and deterministic, but it suffers from a problem in which if any error occurs in the initial alignment and somehow gets propagated to other sequences than it cannot be corrected. However, this type of problem does not exist for iterative methods. In general, iterative methods are much slower in comparison to progressive methods and are used in a place where the best possible alignment is of prime importance and not the computational cost.

Evolutionary algorithms such as the genetic algorithm, which are based on the natural selection processes, are used for implementing iterative methods. Such algorithms have an upper edge with respect to others in the sense that these algorithms are independent for any types of scoring function. This gives an independency that without much alteration to the alignments, different objective functions can easily be tasted. Also, evolutionary algorithms can give low-cost clusters and multi-core processors because of they can be easily parallelize to meets the current trend.

In this study, genetic algorithms (Pengfei et al., 2010[52]) has been considered for experimental analysis. The main advantage of using GA for MSA problem is that it does not requires any particular source of algorithm to solve a given problem. Only, requirement for GA is the fitness function (Dongardive and Abraham, 2012[12]), for necessary analysis and evaluation of solutions. Because GA is an highly implicitly parallel technique therefore, it can be used to solve various large scale and real time problems such as the travelling sales man problem (Zhang and Wong, 1997[71]; Ulder et al., 1991[63]). For a sequences of smaller length it can be possible to do the alignment manually but sequences of larger length requires an algorithm for successful alignment. Progressive alignment technique such as the dynamic programming (DP) suffers from a problem of early convergence or local optima problem and hence cannot be used for alignment of larger sequences. Since, this research work is based on sequences of larger length (see Table 1(Tab. 1)) therefore approaches like GA is considered over DP.

Analyzing the importance of protein sequences in near future (Thompson et al., 2011[61]) provoked the author for considering MSA of protein sequences for this research work. Till date, sequence homology is considered to be the main method for predicting protein structure and function along with their evolutionary history (Kimura, 1980[31]). It has been observed that in the recent years, the tools (Gelly et al., 2011[16]) for MSA of protein sequences has improved. Various literature and related studied have confirmed that the further improvement in protein sequences can only be possible by combining sequence alignment with some know protein structures. A better performance of alignment of protein sequences can be excepted by proper utilizing the phylogenetic relationships among sequences (Cai et al., 2000[7]).

Literature studies (Wong et al., 2000[69]; Taylor, 2000[58]; Razmara et al., 2009[54]; Mott, 2005[41]) says that there are still a number of challenges in aligning protein sequences. First, the misaligned or less aligned locally conserved regions within the sequences are major and foremost challenges in aligning protein sequences. Second, the misalignment of motif which is found in natively disordered regions. Third, the protein sequences which are found in various databases across the globe contain huge amount of alignment error (Loytynoja and Goldman, 2008[36]).

On the basic of literature survey (Devereux et al., 1984[11]; Jagadamba et al., 2011[27]; Nguyen and Yi, 2011[45]; Razmara et al., 2009[54]; Mott, 2005[41]) and in order to test the feasibility of the proposed approach a comparison study were made between the proposed method and some of the existing methods such as the SAGA (Notredame and Higgins, 1996[46]), MSA-GA (Gondro and Kinghorn, 2007[18]), RBT-GA (Taheri and Zomaya, 2009[57]), CLUSTALX (Thompson et al., 1997[59]), CLUSTALW (Thompson et al., 1994[60]), HMMT (Eddy,1995[14]), PRRP (Gotoh, 1996[19]), PILEUP8 (Devereux et al., 1984[11]) and DIALI (Morgenstern et al., 1996[40]) by calculating the corresponding BAliscore. Some of these methods are iterative and some of these are progressive. Each of these methods has their own advantages and disadvantages in terms of speed, time, convergence, robustness and ability to align different lengths sequences etc. All such factors which promoted the author to select these different methods for the experimental study are mentioned in the paragraph that follows.

SAGA, MSA-GA and RBT-GA are the GA based methods. The time complexity of SAGA is larger and are not suffers from the problem of local minima. RBT is an iterative algorithm for sequence alignment using a DP table. CLUSTALW can be seen as an example of progressive approach, and can be used to short out the local optimality problem for the progressive alignment approach. This is the most popular, accurate and practical method in the category of hierarchical methods. The widely used programs for MSA are CLUSTAL W and CLUSTAL X. They are very fast and easy to handle and are capable of aligning datasets of medium sized. The sequences so produced by these methods are of sufficient quality and not requires any manual editing or adjustment. HMMT is based on simulated annealing method. PRRP is a global alignment program which is based on a progressive and iterative approach. This approach is robust. PIMA (Smith and Smith, 1992[56]) uses a local dynamic programming to align only the most conserved motifs. DIANLIGN (Morgenstern et al., 1996[40]) uses a local alignment approach that construct MSA based on a segment to segment comparison rather than residue to residue comparison.

T-Coffee (Notredame et al; 2000[47]) method which was able to make very accurate alignments of very divergent proteins but only for small sets of sequences and therefore not considered for this experimental study. Also this method is often tapped at local minima. It also has a high computational cost with respect to other methods mentioned above. MAFFT (Katoh et al., 2005[29]) is very fast and can align sequences ranging from hundred to thousand. It is quite similar to CLUSTAL when it comes to alignment accuracy. But we have also not considered this method in the proposed research work, as the dataset and the fitness measure used by this algorithm is totally different than those used in this experimental approach.

The rest of the paper is organized as follow. The next section describes the relevant preliminaries on Alignment, Sequence alignment, MSA, GA, BAliBase and PAM Matrix, followed by the proposed approach section which describes the concepts underlying the research work. The experiments setups required in order to validate and observe the results are discussed in the next section. The second last section explains about the detailed results over different datasets. Finally, the concluding section presents the final consideration.

This section provides a detail idea about the basic concept of the related terms used in the paper such as Alignment, Sequence Alignment, Multiple Sequence Alignment, GAP, BAliBase and PAM Matrix.

This section detailed about the proposed approach which is based on various parameters and are described below.

Step 1 : Population initialization x₁,x₂,...,x_n.

Step 2 : Column(N) = 1.2 x n_max. Gaps (-) may be placed in the sequences for proper alignment.

Step 3 : Compute fitness.

Step 4 : Select individuals for genetic operations. Two different genetic operators mainly crossover and mutation is used with probability of 0.8 % and 0.01 %.

Step 5 : Do crossover operation by randomly choosing any one of the defined crossover operator.

Step 6 : Randomly choose and apply all of the defined mutation operator one by one.

Step 7 : Check all the four solution quality, and choose the one who is the best among all four solutions in terms of scores.

Step 8: New population generated and fitness evaluated.

Step 9 : Stop if sufficient solution quality or max search terms reached, which is 50 iteration.

This section gives an overview of the parameters and the systems components used for the experiment.

In this section, the experimental methodology followed in this work is detailed. Moreover, results obtained with the proposed method are presented and discussed.

For all the tests, the different crossover and mutation operators are randomly chosen with equal probability of selection within each generation. To test the proposed approach, the experiments are carried out with different datasets (ref. 1, ref. 2 and ref. 3) of different lengths from the BAliBase database (refer Table 1(Tab. 1)). The author used these datasets for the experimental study because of their performance with other related algorithm, which are gained by referring various literature studies (Devereux et al., 1984[11]; Jagadamba et al., 2011[27]; Nguyen and Yi, 2011[45]; Razmara et al., 2009[54]; Mott, 2005[41]). As stated earlier, for every experiment the alignments were performed with the proposed method and were compared with the methods described in the literature study stated earlier.

For evolution of the proposed approach, the algorithm were executed for 50 independent run (iterations) for 30 datasets (some of all datasets in Table 3(Tab. 3), 4(Tab. 4) and 5(Tab. 5)) and then the best, average and the worst score were calculated. Table 1(Tab. 1) indicates the best, average and the worst score over different datasets with their corresponding BAliscores. As, the fitness score depends upon the level of similarity among the residue in the sequences therefore, the scores can be either positive or negative. Here, one point is to be noted that if the residues among the comparable sequences are similar, then small numbers of gaps (“-”) are needed to make the sequences aligned properly. On the other hand, if the majority of the residues are dissimilar then a large number of gaps are needed for necessary sequence alignment.

To analyze the quality and accuracy of solutions produced by the proposed approach, we have considered a BAliscore, which is an open source program of the BAliBase benchmark. BAliBase scores a solution (multiple sequence alignment) between 0.0 and 1.0. A score of 1.0 indicates that the solution is same or identical to that of manually created reference alignment. Unfortunately, with the proposed approach we are unable to get a score equals to 1(see Tables 3(Tab. 3), 4(Tab. 4) and 5(Tab. 5)). If the score is 0 then it indicates that nothing matches to the reference alignment. This can be observed with some of the datasets in Table 4(Tab. 4) (reference 3).The score between 0 and 1 indicates that some part matches with the reference alignment. The scores which are closer to 1, gives a better alignment for a given dataset. A comparison over different datasets with different methods is being made in Tables 3(Tab. 3), 4(Tab. 4) and 5(Tab. 5). By referring to these tables, we can conclude that the proposed method solution is much more efficient than other methods in terms of scores as indicated in the tables. Figures 8(Fig. 8), 9(Fig. 9), 10(Fig. 10), 11(Fig. 11) and 12(Fig. 12) shows comparative results between the proposed and the other methods discussed in the literature review earlier. Figures 13(Fig. 13), 14(Fig. 14) and 15(Fig. 15) indicates about the average scores comparison among different methods and gives a clear indication about the superiority of the proposed approach over the others.

In order to evaluate the overall performance of the proposed method, the average score of all test cases were evaluated (bottom of Tables 3(Tab. 3), 4(Tab. 4) and 5(Tab. 5)). The average score suggest that the proposed method approach is better among all other methods that are considered. The score is calculated considering the standard BAliBase dataset. The bold faced data`s in the tables indicates the best scores among the methods.

As we all know that the multiple sequence alignment is a known problem in bioinformatics, but still MSA remains a challenging task to explore. The arrangement of molecular sequences within an alignment to find similarities and differences among them is not an easy task, due to the complex size of the sequences and the search space. Because of the ability to handle complex scale problems, genetic algorithm is used as a genuine solution for the multiple sequence alignment problem. In this paper, a novel approach has been developed, which uses genetic algorithm for performing multiple sequence alignment. The motive of the study reported in this paper is to judge the efficiency of the proposed approach by comparing it with different algorithm over standard datasets. In order to evaluate the efficiency and feasibility of the proposed approach, a benchmark datasets from BAliBase 2.0 is considered, because most of the methods discussed in this paper uses BaliBase datasets to access the quality of the multiple sequence alignments. When compared to other methods listed in (Notredame and Higgins,1996[46]; Gondro and Kinghorn, 2007[18]; Taheri and Zomaya, 2009[57]; Thompson et al., 1997[60]; Eddy, 1995[14]; Gotoh, 1996[19]; Devereux et al., 1984[11]; Morgenstern et al., 1996[40]), the proposed method improves the overall quality of the alignment. The experimental result provides a better scope for multiple sequences alignment, as there is an increase in the alignment quality, which can be observed by the scores of different datasets. It was also observed that the proposed method solution gives some unsatisfied results in some test cases. By the above discussions, we can easily conclude that the innovative approach adopted in this paper gives a better and improved result when compared with other methods in most of the testcases.